GeoBurst:*Real.time*Local*Event* Detection*in*Geo.Tagged*Tweet*Streams Chao*Zhang 1 ,*Guangyu*Zhou 1 ,*Quan*Yuan 1 ,*Honglei*Zhuang 1 ,** Yu*Zheng 2 ,*Lance*Kaplan 3 ,*Shaowen*Wang 1 ,*Jiawei*Han 1* 1 UIUC* 2 Microsoft*Research* 3 U.S.*Army*Research*Lab 1
What*is*a*Local*Event? • A*local*events*is*an* unusual&activity *bursted*within*a* local& area *and* specific&duration *while*engaging*a*considerable* number*of*participants.* E.g.,*parade,*riot,*sport*game,*concert,*accident,*disaster. ‣ 2
Local*Event*Detection • Real.time*local*event*detection*is*important*for*various* applications* disaster*monitoring* ‣ crime*alarming* ‣ activity*recommendation* ‣ … ‣ 3
Why*Geo.Tagged*Tweet*Stream? • Real.time*local*event*detection*is*nearly*impossible*years*ago* due*to*the*lack*of*timely*and*reliable*data*sources.** • The*geo.tagged*tweet*stream*brings*new*opportunities*to* this*problem*because*of*its*(1)*sheer*size;*(2)*multi. dimensional*information;*and*(3)*real.time*nature. 4
Our*Goal • Given*the*geo.tagged*tweet*stream,*we*aim*to** detect*all*local*events*in*any*query*time*window*( batch&mode );* ‣ update*the*result*list*in*real*time*as*the*query*window*shifts* ‣ continuously*( online&mode ). query window Q time 5
Challenges 1.Integrate&multiple&types&of&data.& ‣ Location,*time*and*text*have*totally*different*representations.* 2.Extracting&interpretable&events&from&massive&noise.& ‣ Raw*tweets*are*extremely*noisy*and*short.* 3.&On>line&and&real>time&detection.& ‣ To*allow*for*timely*actions,*local*events*should*be*detected*in*real* time. 6
Previous*Studies • Most*existing*event*detection*methods*are*designed*for* detecting* global&events * ‣ They*can*successfully*detect*events*that*are*bursty*in*the*entire* stream;* ‣ But*local*events*are*“bursty”*in*a*small*region*and*involve*a*limited* number*of*tweets.* • A*few*methods*for*local*event*detection*have*been*proposed* They*either*do*not*model*the*correlations*between*keywords;*or* ‣ are*incapable*of*detecting*local*events*in*real*time. 2011 ICWSM. Event detection in twitter. 2012 CIKM. Twevent: segment-based event detection from tweets. 2009 CIKM. Event detection from Flickr data through wavelet-based spatial analysis. 2013 PVLDB. EventTweet: Online localized event detection in the twitter stream. 7
Our*Insight • A*local*event*usually*leads*to*many*related*tweets*around*the* location* (a&geo>topic&cluster) .* • But* a&geo>topic&cluster&is¬&necessarily&a&local&event :* It*may*be*a*routine*activity*in*that*region*(e.g.,*shopping).* ‣ It*may*be*a*global*event*rather*than*a*local*one*(e.g.,*TV*show). ‣ We define a local event as a geo-topic cluster that shows clear spatiotemporal burstiness. 8
Overview*of*GeoBurst • We*propose*GeoBurst,*a*reference.based*method*for*local* event*detection.*It*consists*of*three*key*components:* a&candidate&generator *that*finds*geo.topic*clusters*in*the*query* ‣ time*frame,*and*regard*them*as*candidate*events;* a&ranking&module *that*summarizes*the*routine*activities*in*different* ‣ regions*to*filter*non.event*candidates.* an&updater *that*updates*local*events*in*real*time*as*the*query* ‣ window*shifts. 9
Candidate*Event*Generation • The*candidate*generator*finds*geo.topic*clusters*in*the*query* time*frame*as*candidate*events.* • Geo.topic*cluster:*a*group*of*tweets*that*are*geographically* close*and*semantically*relevant.* • Challenges*for*finding*geo.topic*clusters:* How*to*combine*geographical*and*semantic*similarities?* ‣ How*to*capture*the*correlations*between*different*keywords?* ‣ How*to*cluster*without*knowing*the*number*of*clusters*in*advance? ‣ 10
Candidate*Event*Generation • Intuition:*the*spot*where*the*event*occurs*is*acting*as*a* pivot * that*produces*relevant*tweets*around*it.* • Our*clustering*algorithm*is*based*on:* a*geo.topic*authority*score*for*each*tweet* ‣ an*authority*ascent*process*to*find*authority*maxima*as*pivots ‣ 11
Geo.topic*Authority • A*tweet*gets* an&authority&score *from*neighbor*tweets*where* • the*geographical*impact*is*captured*by*kernel*function;* • the*semantic*impact*is*captured*by*random*walk*on*the*keyword* co.occurrence*graph. music, show music B music semantic A authority geo-impact E impact D shop music, band C Authority can be interpreted as the total amount of energy received from the neighbors. band 12
Pivot • A* pivot *is*an*authority*maximum:*a*prominent*tweet*that*is* surrounded*by*many*relevant*tweets. music, show music B music A E D shop music, band C band 13
Authority*Ascent • Now*the*task*is*to*find*all*the*pivots*in*the*geo.topic*space.* • We*design*an* authority&ascent *process*to*find*all*pivots.* • A&pivot&attracts&similar&tweets *to*form*geo.topic*clusters. neighborhood local pivot d 1 d 2 d 3 neighbor pivot 14
The*Ranking*Module • We*design*the* activity&timeline&structure *to*summarize*the* activities*in*different*spatial*regions*and*time*periods.** • The*summaries*in*the*activity*timeline*serve*as*background* knowledge*to*quantify*the*spatiotemporal*burstiness*of* candidates. snapshot activity timeline time Each snapshot is a set of micro-clusters. Each cluster is an activity summary for a region. 15
The*Ranking*Module • Retrieve*the*snapshots*in*a*reference*window*as*background* knowledge.* • Compute*z.score*for*each*candidate*as*its*ranking*score.* 16
The*Update*Module • In*the*entire*process*of*GeoBurst,*the*most*time.consuming* step*is*pivot*finding.** • How*to*avoid*finding*pivots*from*scratch*as*the*query* window*shifts?* The*key*is*to*maintain*the*local*pivot*for*each*tweet. ‣ neighborhood local pivot d 1 d 2 d 3 neighbor pivot 17
The*Update*Module • We*design*an*updating*strategy*based*on*the*additive* property*of*authority*score:* subtracting*the*contributions*of*outdated*tweets* ‣ emphasizing*the*contributions*of*new*tweets. ‣ neighborhood local pivot d 1 d 2 d 3 neighbor pivot 18
Experimental*Settings • Data:** • NY:*9M*geo.tagged*tweets*in*New*York*during*3*months.* • LA:*8M*geo.tagged*tweets*in*Los*Angeles*during*3*months.* • Task:*80*queries*with*different*durations*(3h,*4h,*5h,*6h),*find* top.5*local*events*in*each*query*window.* • Compared*Method:*EvenTweet*(PVLDB’13),*Wavelet*(CIKM’09)* • Evaluation:*The*crowdsourcing*platform*CrowdFlower* Ask*the*workers*to*judge*whether*the*result*is*a*local*event*or*not. ‣ 19
Illustrative*Cases 20
Precision 21
Running*Time 1. GeoBurst*is*more*efficient*than*the*compared*methods*even*when*in*batch*mode.* 2.The*online*mode*of*GeoBurst*is*more*efficient. 22
Summary • We*study*the*problem*of*detecting*local*events*from*the*geo.tagged* tweet*stream.* • We*proposed*the*GeoBurst*method.* It*first*detects*candidate*events*based*on*authority*ascent,*and*then*ranks*the* ‣ candidates*based*on*background*knowledge.* It*also*features*an*updating*module*to*continuously*monitor*the*stream.* ‣ • Experiments*demonstrate*the*effectiveness*and*efficiency*of*GeoBurst.* • For*future*work,*we*plan*to*extend*GeoBurst*to*handle*the*tweets*that* mention*geo.location*names*but*do*not*have*GPS*information. 23
Thanks! 24
Recommend
More recommend
